Labeling density for 3D training data


Is it required to manually label all nuclei in a 3D training patch for StarDist training? Or is it also possible to do a more sparse labeling and provide some examples in a patch while using more (diverse) patches?


Hi Michiel,

yes you have to annotate all nuclei, including those only partially visible. Here’s some recent information from the developers themselves: [NEUBIAS Academy@Home] Webinar “Introduction to nuclei segmentation with StarDist” + QUESTIONS & ANSWERS. You’ll find tons of other useful information in that post.

As annotating in 3D can be a pain, maybe you can also find useful information in this post where several 3D annotation tools are compared: Comparaison of some tools for 3D dense ground truth annotations

Good luck


1 Like

Thanks Guillaume! That answers my question. Also thanks for the useful links; they contain a lot of info.